Finance Research Letters xxx (xxxx) xxx–xxx
Contents lists available at ScienceDirect
Finance Research Letters journal homepage: www.elsevier.com/locate/frl
Internet search-based investor sentiment and value premium Antti Klemola University of Vaasa, Department of Accounting and Finance, P.O. Box 700, Vaasa FIN-65101, Finland
ARTICLE INFO
ABSTRACT
Keywords: Search-based investor sentiment Internet searches Cross-sectional stock returns Value premium
We study how unexpected change in Internet search-based investor sentiment affects subsequent value premium in the U.S. stock market. For the investor sentiment, we use a sentiment that is based on individual investors’ Internet search activity. We argue that stocks that are considered to be more sensitive to fluctuations in investor sentiment, like financially distressed (proxied by high book-to-market ratio) stocks, should also be more affected by unexpected changes in the sentiment. We find that an unexpected increase in optimism (pessimism) in the sentiment predicts positive (negative) subsequent value premium in the U.S stock market.
JEL: G40
1. Introduction In their seminal work, Baker and Wurgler (2006) show that the investor sentiment does not only affect the aggregate U.S. stock market returns, but they also document a relation between investor sentiment and cross-sectional U.S. stock returns. The authors argue that the effect of sentiment is stronger for stocks that are more difficult to arbitrage and value; for example small, growth, and distressed stocks. Both low (growth) and high (distress) book-to-market ratio stocks are reported to be the most sensitive to the fluctuations in investor sentiment, leading into a U-shaped sensitivity pattern, and the effect of investor sentiment being generally stronger for high book-to-market stocks.1 The authors also note that the effect of investor sentiment on future stock returns is stronger during negative sentiment periods. On a more international level, Baker et al. (2012) and Corredor et al. (2015) also document a closely similar association between investor sentiments and future returns of stocks sorted by their book-to-market ratios. In academic literature, the investor sentiment is usually considered to be either surveys-based (like American Association of Individual Investors and Consumer Confidence) or market-based (like VIX and put-call ratio) or the combination of these two. However, Da et al. (2015) highlight several important arguments why an investor sentiment inferred from Internet search volumes might have an advantage over the previously mentioned and more traditional investor sentiment measurements. First, the market-based sentiment might be the equilibrium outcome of many different economic forces and hence not purely reflect the current investor sentiment. Second, some survey-based sentiments are conducted on too low frequency. Third, the respondents might not answer truthfully in the surveys. Whereas, the Internet search-based investor sentiment also reveals real attitudes rather than just conducting a Gallup poll about it. In recent decade, a new line of research has emerged in the academic literature. These studies especially analyze the impact of Google search volumes on asset prices, where the Google search volumes can be seen as one form of investors’ information retrieval and market attention. Da et al. (2011) find that increase (decrease) in Google search volumes for the stock tickers of Russell 3000 companies predict positive (negative) subsequent returns for the stock in question. Da et al. (2011) also find that an increase in Google search volume for the stock tickers of IPO companies predicts a higher first-day IPO returns. Vozlyublennaia (2014) and
1
E-mail address:
[email protected]. However, they do not find a statistically significant association between lagged investor sentiment and value premium.
https://doi.org/10.1016/j.frl.2019.06.022 Received 18 December 2018; Received in revised form 16 May 2019; Accepted 29 June 2019 1544-6123/ © 2019 Elsevier Inc. All rights reserved.
Please cite this article as: Antti Klemola, Finance Research Letters, https://doi.org/10.1016/j.frl.2019.06.022
Finance Research Letters xxx (xxxx) xxx–xxx
A. Klemola
Klemola et al. (2016) find that information inferred from Google search volumes can also help to predict future aggregate stock market returns. Da et al. (2015) extend the previously mentioned Google studies more into investor sentiment literature by constructing a macrobased Internet search-based investor sentiment from Google search volumes. The authors construct the sentiment by aggregating the Google search volumes for such terminology as recession, unemployment, and bankruptcy.2 Also, Klemola (2018) uses Google search volumes to construct an Internet search-based investor sentiment, but with terminology that is more specifically related to equity market conditions3and terminology used in AAII-survey. Both studies find that their investor sentiments are associated with future near-term aggregate stock market returns. Da et al. (2015) and Klemola (2018) also document a relation between their sentiments and subsequent returns of small-stocks. Klemola (2018) also finds a statistically significant association between the Internet searchbased investor sentiment and size premium. The purpose of this paper is to further study the effect of Internet search-based investor sentiment on stock returns, by utilizing the same Internet search-based investor sentiment as in Klemola (2018), and study its effect on the subsequent value premium in the U.S. stock market. The findings can support the arguments of Da et al. (2015), that the usage of Internet search-based investor sentiment as an alternative investor sentiment measurement is a valid method. As Baker and Wurgler (2006) find, a good investor sentiment should not only effect stock returns on an aggregate level, but it also should have a cross-sectional stock return effect. Thus, this paper aims to strengthen the cross-sectional validation of Internet search-based investor sentiment as an alternative and more modern investor sentiment measurement. The paper contributes to the literature in two separate ways. First, the paper contributes to the studies of Da et al. (2015) and Klemola (2018), by extending the literature of Internet search-based investor sentiment's effect on asset prices to cover also value premium. Second, the paper contributes to studies of Baker and Wurgler (2006), Baker et al. (2012) and Corredor et al. (2015) by studying the effect of investor sentiment on value premium, by utilizing an alternative and more modern investor sentiment measurement. Consistent with our hypothesis, high book-to-market (a proxy for financial distress) stocks are the most affected by unexpected changes in the Internet search-based sentiment. We find that an unexpected increase in optimism (pessimism) in the sentiment predicts positive (negative) value premium for the next week in the U.S. stock market. 2. Data The weekly data used in this study are obtained from multiple sources. The data for Google search volumes are downloaded from Google Trends. The search terms used in this study are bear market and bull market. Furthermore, the popularity of searches is limited to cover only the United States and its finance-related searches. The search volumes are scaled to range from 0 to 100 annually, where zero represents low relative popularity, and 100 represents high relative popularity for the given search terms during the week in question. The data for the returns of 10 different portfolios sorted by their book-to-market ratios are obtained from Kenneth R. French Data Library.4 The choice of control variables is closely similar to Da et al. (2015). For the macroeconomic condition control variable, we use the ADS index developed by Aruoba et al. (2009).5 The ADS contains information on several seasonally-adjusted macroeconomic activities, including weekly initial jobless claims, monthly payroll employment, industrial production, and real domestic product. As a control variable for economic uncertainty, we use the US Economic Policy Uncertainty Index (EPU) as developed by Baker et al. (2016).6 It is based on newspaper coverage frequency of policy-related economic news. As a control variable for equity market uncertainty, we use the US Equity Market Uncertainty Index (EMU).7 Instead of measuring the policy-related economic news (EPU), the EMU measures news related to equity market conditions. We also include the Chicago Board Options Exchange volatility index (VIX)8 as a control variable. In total, the data set consists of 704 weekly observations, starting from the beginning of January 2004 and ending at the end of June 2017. 3. Hypotheses development The methodology of our Internet search-based investor sentiment is based on the investor sentiment develop by Klemola (2018), who use the weekly popularity of Google search terms bull market and bear market or in their difference in popularity (known as a spread). As Peltomäki et al. (2017) and Klemola (2018), we divide our sentiment into two components, expected and unexpected sentiment using AR(1)-process: 2
This Internet search-based investor sentiment is known as FEARS. This Internet search-based investor sentiment is known as Small Investors’ Internet Sentiment (SIIS). 4 Downloaded from http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html 5 Downloaded from https://www.philadelphiafed.org/research-and-data/real-time-center/business-conditions-index 6 Downloaded from http://www.policyuncertainty.com/ 7 Downloaded from http://www.policyuncertainty.com/ 8 Downloaded from Datastream 3
2
Finance Research Letters xxx (xxxx) xxx–xxx
A. Klemola
Sentiment j,t = c j +
Sentiment j,t
E[Sentiment j,t] = c j +
UE[Sentiment j,t] =
1
+
Sentiment j,t
(1)
j,t
(2)
1
(3)
j,t
where E[] is the expected sentiment, and UE[] is the unexpected sentiment. We hypothesize that an unexpected change in the sentiment, UE[Sentiment], represents a shift in noise traders’ beliefs that creates a liquidity shock in the stock market (see, e.g., Campbell et al., 1993). The effect should be stronger for those stocks that are more prone to the behavior of noise traders and are also more difficult to value and arbitrage; like potentially financially distressed (high book-to-market ratio) stocks. We test the effect of unexpected changes in the sentiment to subsequent cross-sectional stocks returns with the following regression model: 4
Ri, t = ci + UE [Sentiment ] j, t
1
+ Ri, t
1
+
h Controlh, t 1
+
h=1
i, t
(4)
where Ri,t is the return of portfolio i consisting of stocks that are sorted by their book-to-market ratios. UE[Sentiment]j,t-1 is one week lagged unexpected change in the Internet search-based investor sentiment inferred from search term j. 4. Empirical analysis Table 1 presents results when the sentiment is inferred from the popularity of the bear market. We document linearly increasing negative relation between the unexpected changes in the sentiment and subsequent cross-sectional stock returns. One standard deviation unexpected increase in the popularity of bear market predicts nine basis points lower return for low book-to-market, and 12 basis points lower return for high book-to-market stocks for the next week. We also document a statistically significant negative crosssectional return spread (value premium) between the high book-to-market and low book-to-market stocks. One standard deviation unexpected increase in the popularity of bear market predicts six basis points lower value premium for the next week. Table 2 presents results when the sentiment is inferred from the popularity of the bull market. We do not document any statistically significant association between the individual portfolios sorted by their book-to-market ratio and unexpected changes in the sentiment. We do however, find that unexpected increase in the popularity of bull market predicts higher value premium for the next week. One standard deviation unexpected increase in the popularity of bull market predicts nine basis points higher value premium for the next week. Table 3 presents results when sentiment is inferred from the difference in popularity of bull market and bear market (the spread). We document a linearly increasing positive association between the unexpected changes in the sentiment and subsequent crosssectional stock returns. One standard deviation unexpected increase in the spread predicts six basis points higher return for low bookto-market stocks and 11 basis points higher return for high book-to-market stocks for the next week. We also document a statistically significant value premium between the high-book-to-market and low-book-to-market stocks. An unexpected increase in the spread predicts 11 basis points higher value premium for the next week. 5. Conclusions We find a statistically significant relation between unexpected changes in the Internet search-based investor sentiment and subsequent value premium in the U.S. stock market. An unexpected increase in optimism (pessimism) in the sentiment predicts positive (negative) value premium for the next week. As Baker and Wurgler (2006) also observe, high book-to-market stocks (a proxy for potential financial distress) generally tend to be more strongly affected by the sentiment. We also observe that when Google search term with a negative meaning, bear market, is used as a sentiment proxy; unexpected changes in the sentiment have a broader cross-sectional effect as we document a linearly increasing negative sensitivity between in the sentiment and future returns of stocks sorted by their book-to-market ratio. This result is consistent with the findings of Baker and Wurgler (2006) and Baker et al. (2012), who report that the sentiment has a stronger effect on future stock returns when the sentiment is on a negative side. The previous result is also consistent with the findings of Tetlock (2007), Tetlock et al. (2008) and García (2013). Tetlock (2007) finds that a high level of media pessimism in the Wall Street Journal predicts negative future market returns. García (2013) finds that the effect of news-based investor sentiment on future stock market returns is stronger during recessions than during expansions. Tetlock et al. (2008) find that individual stock prices tend to react in negative wording in firm-specific news stories. The main contribution of the paper is that unexpected changes in the Internet search-based investor sentiment effects on the subsequent value premium in the U.S. stock market. Thus the sentiment has some cross-sectional effect on future stock returns. This finding further validates the use of the Internet search-based investor sentiment as an alternative and more modern investor sentiment measurement, as it has not the only effect on future aggregate stock markets returns (see, e.g. Da et al., 2015 and Klemola 2018), but it also has cross-sectional effect on future stock returns (see, e.g. Baker and Wurgler 2006). The main findings also contribute to studies of Baker and Wurgler (2006), Baker et al. (2012) and Corredor et al. (2015) by further documenting the effect of investor sentiment on value premium, by utilizing the Internet search-based investor sentiment instead. 3
4
−0.008 (−0.036) −0.010*** (−2.585) −0.062 (−1.155) 0.362 (1.608) −0.003 (−0.942) 0.000 (0.295) 0.022 (1.431) 0.016 2.84 703
−0.079 (−0.385) −0.012*** (−2.815) −0.094* (−1.951) 0.408* (1.937) −0.002 (−0.751) 0.001 (0.523) 0.022 (1.503) 0.024 3.87 703
2 −0.012 (−0.056) −0.014*** (−3.429) −0.082* (−1.751) 0.430** (1.987) −0.003 (−1.024) 0.000 (0.326) 0.025 (1.555) 0.029 4.52 703
3 −0.137 (−0.577) −0.015*** (−3.482) −0.059 (−1.122) 0.538** (2.256) −0.003 (−0.831) 0.001 (0.325) 0.030* (1.795) 0.027 4.21 703
4 0.055 (0.205) −0.016*** (−3.231) −0.101* (−1.949) 0.413 (1.645) −0.003 (−0.924) 0.000 (0.225) 0.021 (1.190) 0.028 4.38 703
5 −0.030 (−0.112) −0.017*** (−3.657) −0.092* (−1.803) 0.476** (2.005) −0.003 (−0.806) 0.001 (0.437) 0.024 (1.349) 0.030 4.67 703
6 −0.114 (−0.423) −0.016*** (−3.172) −0.072 (−1.257) 0.665** (2.247) −0.003 (−0.633) 0.001 (0.626) 0.025 (1.346) 0.031 4.76 703
7 0.374 (0.988) −0.018*** (−3.352) −0.094 (−1.409) 0.387 (1.206) −0.004 (−0.801) 0.000 (0.083) 0.005 (0.223) 0.024 3.94 703
8
0.264 (0.726) −0.020*** (−3.461) −0.110** (−2.281) 0.245 (0.834) −0.005 (−1.115) 0.001 (0.841) 0.006 (0.264) 0.024 3.89 703
9
0.273 (0.572) −0.018*** (−2.612) −0.042 (−0.758) 0.353 (0.808) −0.004 (−0.805) 0.001 (0.348) 0.007 (0.238) 0.007 1.87 703
10
0.303 (0.890) −0.008* (−1.737) −0.039 (−0.602) −0.016 (−0.046) −0.001 (−0.441) 0.001 (0.293) −0.017 (−0.917) −0.000 0.99 703
10–1
Notes: This table shows estimated coefficients for Eq. (4), where the sentiment is inferred from the popularity of Google search term bear market. Ten different stock portfolios are sorted by their book-tomarket ratio from lowest (1) to highest (10). 10–1 is a long-short portfolio representing a value premium, formed by being long on top decile and short on low decile of stocks sorted by their book-tomarket ratio. T-stats are reported in parentheses and *, **, *** refer to statistical significance at 0.1, 0.05, 0.01 level.
R2 F-stat Obs
VIXt-1
EPUt-1
EMUt-1
ADSt-1
Rt-1
UE[SIIS]t-1
C
1
Table 1 An unexpected change in Bear Market popularity and future cross-sectional stock returns.
A. Klemola
Finance Research Letters xxx (xxxx) xxx–xxx
Finance Research Letters xxx (xxxx) xxx–xxx
A. Klemola
Table 2 An unexpected change in Bull Market popularity and future cross-sectional stock returns.
C UE[SIIS]t-1 Rt-1 ADSt-1 EMUt-1 EPUt-1 VIXt-1 R2 F-stat Obs
1
2
3
4
5
6
7
8
9
10
10–1
−0.011 (−0.049) 0.000 (0.056) −0.056 (−1.052) 0.341 (1.550) −0.003 (−0.986) 0.001 (0.454) 0.021 (1.380) 0.008 1.92 703
−0.076 (−0.372) 0.002 (0.429) −0.085* (−1.771) 0.383* (1.859) −0.003 (−0.805) 0.001 (0.677) 0.021 (1.409) 0.014 2.68 703
−0.010 (−0.046) 0.001 (0.320) −0.075 (−1.625) 0.400* (1.880) −0.003 (−1.088) 0.001 (0.526) 0.023 (1.443) 0.015 2.82 703
−0.127 (−0.544) 0.003 (0.620) −0.051 (−0.967) 0.505** (2.153) −0.003 (−0.903) 0.001 (0.516) 0.028* (1.661) 0.014 2.66 703
0.066 (0.250) 0.004 (0.664) −0.093* (−1.800) 0.378 (1.527) −0.004 (−0.995) 0.001 (0.422) 0.019 (1.063) 0.015 2.74 703
−0.018 (−0.068) 0.004 (0.693) −0.084* (−1.680) 0.438* (1.880) −0.003 (−0.878) 0.001 (0.657) 0.021 (1.204) 0.015 2.76 703
−0.103 (−0.394) 0.003 (0.595) −0.065 (−1.142) 0.628** (2.173) −0.003 (−0.702) 0.001 (0.830) 0.022 (1.238) 0.019 3.27 703
0.403 (1.075) 0.006 (0.959) −0.092 (−1.357) 0.343 (1.079) −0.004 (−0.867) 0.001 (0.313) 0.002 (0.067) 0.014 2.60 703
0.282 (0.785) 0.004 (0.615) −0.105** (−2.181) 0.200 (0.689) −0.005 (−1.185) 0.002 (1.107) 0.003 (0.121) 0.011 2.25 703
0.326 (0.689) 0.013 (1.498) −0.038 (−0.700) 0.308 (0.713) −0.005 (−0.877) 0.001 (0.513) 0.002 (0.063) 0.004 1.44 703
0.359 (1.047) 0.012** (2.421) −0.039 (−0.602) −0.041 (−0.121) −0.002 (−0.525) 0.001 (0.387) −0.021 (−1.110) 0.005 1.55 703
Notes: This table shows estimated coefficients for Eq. (4), where the sentiment is inferred from the popularity of Google search term bull market. Ten different stock portfolios are sorted by their book-to-market ratio from lowest (1) to highest (10). 10–1 is a long-short portfolio representing a value premium, formed by being long on top decile and short on low decile of stocks sorted by their book-to-market ratio. T-stats are reported in parentheses and *, **, *** refer to statistical significance at 0.1, 0.05, 0.01 level. Table 3 An unexpected change in the spread and future cross-sectional stock returns.
C UE[SIIS]t-1 Rt-1 ADSt-1 EMUt-1 EPUt-1 VIXt-1 R2 F-stat Obs
1
2
3
4
5
6
7
8
9
10
10–1
0.010 (0.044) 0.006* (1.871) −0.056 (−1.067) 0.353 (1.572) −0.003 (−0.978) 0.000 (0.336) 0.021 (1.370) 0.012 2.39 703
−0.056 (−0.277) 0.007** (2.255) −0.086* (−1.817) 0.400* (1.902) −0.003 (−0.790) 0.001 (0.546) 0.021 (1.438) 0.021 3.47 703
0.015 (0.070) 0.008*** (2.775) −0.076* (−1.649) 0.419* (1.939) −0.003 (−1.075) 0.001 (0.365) 0.024 (1.466) 0.024 3.89 703
−0.102 (−0.431) 0.010*** (3.004) −0.054 (−1.012) 0.528** (2.221) −0.003 (−0.880) 0.001 (0.348) 0.028* (1.698) 0.024 3.84 703
0.092 (0.341) 0.011*** (2.745) −0.095* (−1.836) 0.404 (1.608) −0.003 (−0.974) 0.000 (0.247) 0.019 (1.092) 0.025 3.98 703
0.009 (0.035) 0.012*** (3.040) −0.084* (−1.690) 0.465** (1.969) −0.003 (−0.856) 0.001 (0.463) 0.022 (1.253) 0.027 4.22 703
−0.077 (−0.287) 0.011*** (2.658) −0.066 (−1.146) 0.653** (2.211) −0.003 (−0.681) 0.001 (0.655) 0.023 (1.254) 0.028 4.34 703
0.423 (1.101) 0.013*** (2.776) −0.091 (−1.349) 0.376 (1.182) −0.004 (−0.843) 0.000 (0.101) 0.003 (0.123) 0.023 3.77 703
0.314 (0.862) 0.014*** (3.033) −0.106** (−2.214) 0.233 (0.801) −0.005 (−1.163) 0.002 (0.890) 0.004 (0.156) 0.021 3.52 703
0.332 (0.691) 0.017*** (3.036) −0.038 (−0.693) 0.354 (0.816) −0.004 (−0.838) 0.001 (0.313) 0.004 (0.161) 0.011 2.32 703
0.344 (1.003) 0.011*** (3.086) −0.036 (−0.555) −0.007 (−0.020) −0.001 (−0.457) 0.000 (0.214) −0.018 (−0.972) 0.008 1.96 703
Notes: This table shows estimated coefficients for Eq. (4), where the sentiment is inferred from the spread between the popularities of Google search term bear market and bull market. Ten different stock portfolios are sorted by their book-to-market ratio from lowest (1) to highest (10). 10–1 is a long-short portfolio representing a value premium, formed by being long on top decile and short on low decile of stocks sorted by their book-tomarket ratio. T-stats are reported in parentheses and *, **, *** refer to statistical significance at 0.1, 0.05, 0.01 level.
Acknowledgments The author thanks the discussant and other participants at Graduate School of Finance Winter Workshop. The author would also like to thank the Nordea Pankin Säätiö for financial support. References Aruoba, S., Diebold, F., Scotti, C., 2009. Real-time measurement of business conditions. J. Bus. Econ. 27 (4), 417–427. Baker, M., Wurgler, J., 2006. Investor sentiment and the cross-section of stock returns. J. Finance 61 (4), 1645–1680. Baker, M., Wurgler, J., Yuan, Y., 2012. Global, local, and contagious investor sentiment. J. Financ. Econ. 104 (2), 272–287. Baker, S., Bloom, N., Davis, S., 2016. Measuring economic policy uncertainty. Q. J. Econ. 131 (4), 1593–1636. Campbell, J., Grossman, S., Wang, J., 1993. Trading volume and serial correlation in stock returns. Q. J. Econ. 108 (4), 905–939. Corredor, P., Ferrer, E., Santamaria, R., 2015. The impact of investor sentiment on stock returns in emerging markets: the case of central european markets. East. Eur. Econ. 53 (4), 328–355. Da, Z., Engelberg, J., Gao, P., 2011. In search of attention. J. Finance 66 (5), 1461–1499. Da, Z., Engelberg, J., Gao, P., 2015. The sum of all FEARS investor sentiment and asset prices. Rev. Financ. Stud. 28 (1), 1–32.
5
Finance Research Letters xxx (xxxx) xxx–xxx
A. Klemola
García, D., 2013. Sentiment during recessions. J Finance 68 (3), 1267–1300. Klemola, A., Nikkinen, J., Peltomäki, J., 2016. Changes in Investors’ market attention and near-term stock market returns. J. Behav. Finance 17 (1), 18–30. Klemola, A., 2018. Small Investors’ internet sentiment and return predictability. Rev. Behav. Finance In press. Peltomäki, J., Graham, M., Hasselgren, A., 2017. Investor attention to market categories and market volatility: the case of emerging markets. Res. Int. Bus. Finance 44 (April), 532–546. Tetlock, P., 2007. Giving content to investor Sentiment: the role of media in the stock market. J. Finance 62 (3), 1139–1168. Tetlock, P., Saar-Tsechansky, M., Macskassy, S., 2008. More than words: quantifying language to measure Firms’ fundamentals. J. Finance 63 (3), 1437–1467. Vozlyublennaia, N., 2014. Investor attention, index performance, and return predictability. J. Bank. Finance 41 (April), 17–35.
6